Last week somebody rebuilt a working website using a swarm of cheap AI agents bossed around by a smarter one, and the whole thing cost about eight dollars —  sandwich and a coffee. But the part that lit up my corner of the internet wasn’t the website. It was that the agents caught their own lies.

One of them went out and collected a couple hundred quotes for the site and reported back that every last one was verified, checked, gold. Another agent — a suspicious little thing that had been told to ignore the first one’s homework and redo it from scratch — went and compared each quote against the actual source. More than a dozen were wrong. Invented. Stitched together from real fragments into sentences nobody had ever said. The checker flagged them, sent them back, and they got fixed. No human touched it. Total additional cost: a rounding error.

The reaction was what you’d expect. Hallucination is solved. It’s handled structurally now. It’s just a recipe — anybody can do this.

And here’s the thing that made me put down my coffee: they’re right. That’s the interesting part. They are completely, verifiably right, and they have drawn precisely the wrong conclusion from being right.

Let me tell you what I think actually happened last week. Then, because a thesis you haven’t tried to strangle in its crib isn’t worth printing, let me spend a few hundred words trying to prove myself wrong.

The thesis, in one sentence

When everybody can check, “we checked” stops meaning anything — and the value doesn’t disappear, it moves to whoever can prove it.

That’s it. That’s the whole argument. Cheap, universal self-checking doesn’t make verification worthless. It makes verification worthless and makes proof of verification more valuable than it has ever been. Those are two different goods, and we are about to spend the next few years confusing them.

I know this because we have run this exact experiment before. Several times. It always ends the same way, and it has never once ended the way the eight-dollar-website people think it’s going to end.

We have done this before

Once upon a time, a company’s books were whatever the company said they were. You wanted to know if a firm was sound, you asked the firm, and the firm told you it was doing marvelously, thank you. This worked exactly as well as you’d imagine. Victorian England ran on railway shares and joint-stock companies that turned out, with some regularity, to be elaborate works of fiction. America capped the genre in 1929.

Now — anybody could add up a ledger. Arithmetic was not a scarce resource. Double-entry bookkeeping had been sitting in print since a Franciscan friar wrote it down in 1494. The capability to check the numbers was cheap and universal and had been for four centuries. And it counted for nothing, because the party doing the adding was the party with everything to gain from the total.

So the market invented a thing. Not new arithmetic — new arithmetic-with-a-signature-on-it. An independent audit, performed by someone who didn’t work for you, who put their name and their liability on an opinion that your books were what you claimed. After 1933 and 1934 we made it the law for anyone who wanted to sell shares to the public. And a strange thing happened to the value: the signature on the auditor’s opinion became worth more than the numbers underneath it. The numbers were free. Everybody had the numbers. What cost money — what still costs a fortune — was the independent, accountable, provable attestation that the numbers were true.

This is not a one-off. It is the most reliable pattern in the history of commerce, and once you see it you can’t stop seeing it.

John Moody couldn’t tell which railroad bonds were safe, and neither could anybody else, so in 1909 he started publishing ratings — a trusted outsider’s letter grade — and built an institution that outlived every railroad he graded. Anyone could read a balance sheet. What you couldn’t do was vouch.

By 1894 American cities were wiring themselves for electricity and quietly burning themselves down, so a young electrical inspector founded a laboratory to test the equipment and stamp the safe ones with a little circled UL. Anyone could wire a lamp. What you couldn’t do was promise a nervous insurer it wouldn’t kill somebody.

Go back further and it’s a notary — a person whose entire job, for two thousand years, has been to be a trusted third party who watches you sign and swears you signed. Come forward and it’s the little padlock in your browser, which exists because somewhere a certificate authority you’ve never heard of vouched that the website is who it says it is. Anyone can make a web page. What you can’t do, standing there yourself, is prove it’s not a trap.

The pattern, every single time, in four beats: a capability gets cheap and universal; self-attestation becomes worthless because it’s cheap and universal; a trusted, independent third party steps into the gap; and the proof ends up worth more than the thing being proved. The calculator goes to zero. The audit becomes an industry.

We are, right now, at beat one for artificial intelligence. The capability to check an AI’s work is going cheap and universal in front of our eyes — eight dollars, an afternoon, a recipe. Which means, if history is any guide at all, we are standing at the exact threshold where the proof is about to become the whole game.

Now let me try to break it

I don’t trust arguments this tidy, and neither should you. History rhyming is not history repeating, and “a trusted third party always wins” is the kind of thing that sounds like wisdom and behaves like a horoscope. So here are the three ways I can see this time genuinely being different. I take them seriously. You should too.

One: maybe it commoditizes all the way down. Maybe AI verification gets so cheap, so built-in, so ambient that it dissolves into the plumbing and no third party can carve out a slice worth paying for. The checking becomes like spell-check — everywhere, free, and nobody’s business model. That’s a real possibility, and it’s the strongest objection, because the eight-dollar demo is exactly what total commoditization looks like on day one.

Two: maybe the platforms become their own auditors. The biggest AI companies are not fools. They can build the checking into their own products and stamp their own work “verified.” Why would a market pay an independent when the incumbent offers it for free, integrated, with one throat to choke? Perhaps the attestor and the attested merge, and the independent third party never gets born at all.

Three: maybe regulation flattens it. Government could simply mandate a checklist — “thou shalt verify” — turn it into a compliance box every vendor ticks, and reduce the whole thing to paperwork. No premium, no institution, just another line in a form nobody reads.

Those are good objections. If I were sitting across a table trying to poke holes in me, those are the three I’d use. Let me take them in reverse, because they get easier to answer as you go.

Regulation flattening the market is exactly backward, and we have the receipts. When Congress passed Sarbanes-Oxley after Enron, it did not turn auditing into a commodity checkbox. It set off the greatest hiring boom in the history of the accounting profession. Mandating verification doesn’t kill the verifier — it conscripts the whole market into needing one. A rule that says “you must prove it” is a rule that creates permanent, price-insensitive demand for whoever can do the proving. Regulation is the auditor’s best friend and always has been.

The platform-as-its-own-auditor objection is the one that feels strongest and is actually the weakest, because it fails on the single word the entire edifice is built on: independence. You cannot audit yourself. It is not a technical limitation, it is a logical one, and every regulated industry on earth already knows it in its bones. A bank’s regulator will not accept “our vendor checked its own work.” A court will not accept “the company that made the statement also certified it.” And we know precisely what happens when the attestor isn’t independent, because we watched it: Arthur Andersen didn’t collapse in 2002 because it couldn’t do the arithmetic. It collapsed because it audited Enron and consulted for Enron, and the moment the checker and the checked shared a paycheck, the signature became worthless — and took an eighty-nine-year-old firm and its clients down with it. The platforms can build wonderful checking. What they structurally cannot build is independence from themselves. That gap doesn’t close with a better model. It’s not a model problem.

Which leaves the first objection, the real one — total commoditization — and here is where I finally climb down off the fence. The mistake in “it’ll commoditize all the way down” is that it treats the calculator and the audit as the same product arriving at the same time. They are not. The commoditization of the capability is not the enemy of the attestation — it is the precondition for it. Cheap arithmetic didn’t prevent the audit industry; it created it, by making self-reported numbers worthless enough that the market went looking for someone independent to sign them. Spell-check going free didn’t end proofreading at the places where being wrong is catastrophic — publishers, law firms, the FDA label on your prescription. The cheaper and more universal the checking gets, the more obvious it becomes that “we checked” is a thing anyone can say and therefore a thing no one can be trusted on — and the more valuable it becomes to be the one party who can hand a regulator, a court, or a nervous CEO something better than a promise. A receipt. An opinion. A signature with liability attached.

Cheap checking doesn’t drain the pool. It fills it, and then it makes everyone thirsty for the one glass of water nobody’s allowed to pour for themselves.

Where that leaves us

So here’s where I’ve landed, having tried in good faith to land somewhere else.

Everyone racing to make AI check its own work is building the calculator. It’s genuinely impressive, it’s getting cheaper by the week, and it is going to be worth approximately nothing — not because it doesn’t work, but because it works for everybody, which in the trust business is the same as not working at all. The eight-dollar website was not a glimpse of the finish line. It was the starting gun.

The prize — the durable, defensible, boring-in-the-way-that-mints-money prize — is the audit. The independent, provable, accountable verification of record that regulated industries can rely on and, more importantly, defend. Not a model that says “trust me, I checked.” A layer that hands you a signed opinion you can put in front of the people who will hold you responsible when it’s wrong.

We are going to spend the next couple of years watching very smart people confuse those two things. They’ll keep announcing that hallucination is solved, and they’ll keep being technically correct and strategically lost. And somewhere in the middle of all that noise, quietly, the way it happened with the ledger and the bond and the wire and the padlock, somebody is going to build the thing that actually matters.

I’ve been watching this industry long enough to have seen the pattern play out in four other centuries and at least three other technologies. I’d bet the coffee and the sandwich it plays out again.

The numbers are free. They’ve always been free. It’s the signature that costs.